The field of diffusion models is rapidly advancing, with a focus on improving efficiency, quality, and applicability to various tasks. Recent developments have led to the creation of novel frameworks, such as Shortcut Flow Matching for Speech Enhancement, which enables high-quality synthesis in just a few steps using deterministic ordinary differential equation solvers. Additionally, techniques like Discrete Guidance Matching and Stage-wise Dynamics of Classifier-Free Guidance have been proposed to improve the sampling efficiency and quality of diffusion models. Noteworthy papers in this area include 'Shortcut Flow Matching for Speech Enhancement', which achieves a real-time factor of 0.013 on a consumer GPU while delivering perceptual quality comparable to a strong diffusion baseline, and 'HiGS: History-Guided Sampling for Plug-and-Play Enhancement of Diffusion Models', which consistently improves image quality across diverse models and architectures.
Advancements in Diffusion Models for Image and Speech Enhancement
Sources
Shortcut Flow Matching for Speech Enhancement: Step-Invariant flows via single stage training
Discrete Guidance Matching: Exact Guidance for Discrete Flow Matching
Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models
HiGS: History-Guided Sampling for Plug-and-Play Enhancement of Diffusion Models
LucidFlux: Caption-Free Universal Image Restoration via a Large-Scale Diffusion Transformer
Diffusion Models are Kelly Gamblers
SAIP: A Plug-and-Play Scale-adaptive Module in Diffusion-based Inverse Problems
RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement
GLASS Flows: Transition Sampling for Alignment of Flow and Diffusion Models
TR2-D2: Tree Search Guided Trajectory-Aware Fine-Tuning for Discrete Diffusion
ART-VITON: Measurement-Guided Latent Diffusion for Artifact-Free Virtual Try-On
EVODiff: Entropy-aware Variance Optimized Diffusion Inference
Diffusion Alignment as Variational Expectation-Maximization
Learn to Guide Your Diffusion Model
Temporal Score Rescaling for Temperature Sampling in Diffusion and Flow Models
FideDiff: Efficient Diffusion Model for High-Fidelity Image Motion Deblurring
Test-Time Anchoring for Discrete Diffusion Posterior Sampling